Lorenzo Rimoldini1, Laurent Eyer2, Nami Mowlavi2, Dafydd W. Evans3, Krzysztof Nienartowicz4, Berry Holl1, Marc Audard1, Leanne P. Guy1, Gr´egory Jevardat de Fombelle4, Isabelle Lecoeur-Ta¨ıbi1, Olivier Marchal1, Gisella Clementini5, Vincenzo Ripepi6, Alessia Garofalo7·5, Roberto Molinaro6, Tatiana Muraveva5, Ennio Poretti8, L´aszl´o Moln´ar9, Emese Plachy9,
Aron Juh´´ asz9·10, L´aszl´o Szabados9, Joris De Ridder11, Sara Regibo11, Luis Manuel Sarro Baro12and Mauro L´opez del Fresno13
1. Department of Astronomy, University of Geneva, Chemin d’Ecogia 16, CH-1290 Versoix, Switzerland
2. Department of Astronomy, University of Geneva, Chemin des Maillettes 51, CH-1290 Versoix, Switzerland
3. Institute of Astronomy, University of Cambridge, Madingley Road, Cambridge CB3 0HA, United Kingdom
4. SixSq, Rue du Bois-du-Lan 8, CH-1217 Meyrin, Switzerland 5. INAF - Osservatorio Astronomico di Bologna, Via Gobetti 93/3,
I-40129 Bologna, Italy
6. INAF - Osservatorio Astronomico di Capodimonte, Via Moiariello 16, I-80131 Napoli, Italy
7. Department of Physics and Astronomy, University of Bologna, Via Gobetti 93/2, I-40129 Bologna, Italy
8. INAF - Osservatorio Astronomico di Brera, Via E. Bianchi 46, I-23807 Merate, Italy 9. Konkoly Observatory, Research Centre for Astronomy & Earth Sciences, Hungarian
Academy of Sciences, Konkoly Thege Mikl´os ´ut 15-17, H-1121 Budapest, Hungary 10. E¨otv¨os Lor´and University, Egyetem t´er 1-3, H-1053 Budapest, Hungary
11. Institute of Astronomy, KU Leuven, Celestijnenlaan 200D, B-3001 Leuven, Belgium 12. Departamento Inteligencia Artificial, UNED, Calle Juan del Rosal 16,
E-28040 Madrid, Spain
13. Departamento de Astrof´ısica, Centro de Astrobiolog´ıa (INTA-CSIC), PO Box 78, E-28691 Villanueva de la Ca˜nada, Spain
The secondGaia data release is expected to contain data products from about 22 months of observation. Based on these data, we aim to provide an advance publication of a full-skyGaia map of RR Lyrae stars. Although comprehensive, these data still contain a significant fraction of sources which are insufficiently sampled for Fourier series decomposition of the periodic light variations. The challenges in the identification of RR Lyrae candidates with (much) fewer than 20 field-of-view transits are described. General considerations of the results, their limitations, and interpretation are presented together with prospects for improvement in subsequentGaiadata releases.
1 Introduction
RR Lyrae stars are particularly useful variable objects as they combine ease of detection (due to light variations of up to about two magnitudes and periods typically
less than about one day) with the benefits of being standard candles, making studies of structures and distance estimations possible within and beyond our Galaxy.
A low number of observations can suffice in the identification of RR Lyrae stars (Ivezi´c et al., 2000; Hernitschek et al., 2016), so we searched for these objects in the first 22 months of data from the Gaia mission (Gaia Collaboration et al., 2016a), with the hope that the community will benefit of an advance publication of all-sky RR Lyrae candidates inGaia Data Release 2 (DR2). Efforts to identify RR Lyrae stars inGaia DR1 data from sources without time series have already appeared in the literature (e.g., Belokurov et al., 2017; Iorio et al., 2018), indicating that an early publication and exploitation of selectedGaia time series can better address current challenges. Herein, the term ‘observation’ denotes a single field-of-view transit of a source along theGaia focal plane in theGband (typically a combination of 8 or 9 astrometric field CCDs).
2 Method
The minimum number of observations necessary for reliable Fourier series decom- position of the periodic light variations depends on several factors, like the time sampling, the range of fundamental frequencies of the light curve, the number of harmonics needed to characterize the time series, and the level of noise. In the case ofGaia DR1, at least 20 observations were required for most of the published RR Lyrae stars (only in a few cases the number of observations was as low as 12−15). ForGaia DR2 candidates, more than half of the objects have less than 20 observations, with the highest peak of the distribution around 13−14 observations.
In order to recognize RR Lyrae stars, machine-learning models are built with feature-based semi-supervised classification techniques. For an unbiased identifica- tion of candidates associated with high and low (at least two) numbers of obser- vations, classifiers are trained with values of statistical nature without resorting to Fourier modeling parameters. The classification of sources with at least 20 ob- servations, covering (non-uniformly) about half of the sky and including Fourier parameters, is performed by an independent pipeline run, and the results from both approaches will be published inGaia DR2.
Crossmatch with Literature. The crossmatch ofGaiasources with objects from the literature is fundamental to the construction of a realistic training set and the subsequent validation of results. Known RR Lyrae stars are extracted from the following surveys: ASAS (Pojmanski, 1997), Catalina (Drake et al., 2009), Gaia (Gaia Collaboration et al., 2016b; Eyer et al., 2017; Clementini et al., 2016),Hippar- cos (European Space Agency, 1997; Perryman et al., 1997), LINEAR (Stokes et al., 2000), NSVS (Wo´zniak et al., 2004), OGLE-IV (Udalski et al., 2015), Pan-STARRS1 (Chambers et al., 2016), and SDSS (Abazajian et al., 2009), in addition to a selection of globular clusters and ultra-faint dwarf spheroidal galaxies.
Classification Attributes. About 150 attributes are defined to characterize dif- ferent aspects of the time series and a subset of 40 attributes is selected according to the attribute usefulness as perceived by the classifier (e.g., Guyon & Elisseeff, 2003).
The classification attributes that are found particularly relevant to the identification of RR Lyrae stars in the Gaia data include (i) the amplitude of light variations,
(ii) the skewness of the distribution of magnitudes, (iii) the typical magnitude range (and its fraction with respect to the full time series) when restricted to time series segments of one or half a day, and (iv) the interquartile range of the distribution of absolute values of magnitude changes per unit time between successive observations.
Classification Model. Crossmatched objects unavoidably suffer from the selec- tion biases present in the originating catalogs, due to the peculiarities and limitations of each survey (sky coverage, sampling, photometric bands, sensitivity, etc.). Re- sampling is used to alleviate density peaks in the distributions of some parameters, and a semi-supervised approach is applied to the training set of certain classes (con- stant stars, fundamental and first-overtone RR Lyrae stars, and Mira variables) by selecting results from a previous classification run to improve the training set rep- resentation in the sky and in magnitude distribution. The selection of unlabeled sources to add to the training set depends on the parameter gap(s) to fill and on the classification probability (sufficiently high to limit chances of contamination, but not too high in order to allow the discovery of new information). In order to ensure confidence in the new objects, unlabeled sources are verified through basic statistical filters as a function of variability type before adding them to the training set.
Random Forest (Breiman, 2001) classifiers are implemented in a multi-stage set- ting which combines a set of dedicated classifiers to solve simpler problems with fewer attributes, making data less diluted and thus better represented in attribute space. Multi-stage classification avoids also the potentially counter-productive com- petition which could occur in the identification of major classes and sub-classes at the same level. Our multi-stage classifier includes five dedicated classifiers to sepa- rate (i) constant objects from low-amplitude variables and other variables, (ii) low- amplitude variability types, (iii) other variability classes, (iv) RR Lyrae stars into sub-types (fundamental, first overtone, double mode and anomalous double mode), and (v) Cepheids into sub-types (anomalous, classical, type-II). Classifications from (ii) and (iii) depend on the results of (i), while those from (iv) and (v) depend on (iii).
A few other variability types (besides RR Lyrae stars) from the all-sky clas- sification will also be published in Gaia DR2: Cepheids, high-amplitude δ Scuti and SX Phoenicis stars, and long period (Mira and semi-regular) variables. Addi- tional classes are included (although not published) in different classification stages to reduce the contamination of the targeted classes. The average completeness and contamination rates (per type) of the classifier that distinguishes the main variabil- ity types published inGaia DR2 are shown in Fig. 1 and are applicable to sources with attribute distributions similar to those of the training set.
3 Results
The classification results are associated with a score to quantify the reliability of the candidates. The number of true and false positives among the RR Lyrae clas- sifications is assessed statistically in terms of completeness and contamination rates as a function of classification score, magnitude, and number of observations. These rates vary also as a function of extinction, reddening, crowding, and possibly other parameters, which need to be considered (depending on the context of the analysis) for a correct interpretation of the identified candidates.
As expected, the distribution in the sky reveals the absence of reliable candidates
3 1 4 11 4 1 3 74 2 6 1 90 1
1 97 1
98 1
1 94 1 3 1 1 81 5 2 11 1 74 4 10 1 6 6 1 68 1 5 2 2 19 4
50 50
RS RRAB_RRC_RRD_ARRD QSO MIRA_SR ECL DSCT_SXPHE CV CEP_ACEP_T2CEP BLAP
%
%
%
%
%
%
%
%
%
BLAP CEP_ACEP_T2CEP CV DSCT_SXPHE ECL MIRA_SR QSO RRAB_RRC_RRD_ARRD RS
500 4108 1996 1945 3861 1326 518 711 10
# Obj. / Class
Contamination − 11 10 13 12 3 7 11 24%
Fig. 1: Confusion matrix of the classifier addressing the main classes published inGaiaDR2:
training-set objects (in rows) and their numbers (on the left-hand side) versus pre- dictions from the classifier (in columns), where percentages along the diagonal repre- sent the completeness rates. Trained classes include BLAP (blue large amplitude pul- sators), CEP ACEP T2CEP (classical, anomalous, and type-II Cepheids), CV (cataclysmic variables), DSCT SXPHE (δ Scuti and SX Phoenicis stars), ECL (eclipsing binaries), MIRA SR (Mira and semi-regular variables), QSO (quasars), RRAB RRC RRD ARRD (fundamental, first-overtone, double-mode, and anomalous double-mode RR Lyrae stars), and RS (RS Canum Venaticorum-type binary systems). Candidates of the under- represented BLAP class are merged with those of the DSCT SXPHE class inGaia DR2.
in the most reddened and extinguished regions of the Galactic disc. Many known objects such as globular clusters, dwarf spheroidal galaxies, and the Sagittarius tidal streams are easily distinguishable in a simple sky map of RR Lyrae candidates.
Multiple independent validations of the results are performed for each of the published variability types. In the case of RR Lyrae classifications, such validations include: (i) known objects from the literature to assess the statistical quality of re- sults and reduce obvious contaminants; (ii) sources detected in theKepler/K2 fields (characterized by finely sampled light curves) to search for counterparts ofGaia can- didates and estimate the fraction of missed identifications; (iii) the pipeline module dedicated to the confirmation of RR Lyrae stars, usually executed with at least 20 observations, but now extended to include sources with as few as 12 observations.
Full details of the results and comparisons with recent works in the literature will be presented in the documentation and articles accompanyingGaia DR2.
4 Prospects
The data in Gaia DR2 will include astrometric information, time series in theG, GBP,GRPbands, and statistical parameters of the published RR Lyrae candidates.
More detailed information (such as period, modeling parameters, metallicity and
extinction) will be available for a subset of the sources with at least 12 observa- tions that could be confirmed by the RR Lyrae verification module with detailed Fourier modeling. FutureGaia data releases are expected to provide identifications of RR Lyrae stars with higher completeness and lower contamination rates, due to more observations, improved astrometry, radial velocities, spectra, and astrophysical parameters (such as temperature and extinction), some of which might already be available inGaia DR2 (although not in time to be used in this classification).
Acknowledgements. This work made use of data from the ESA space mission Gaia, pro- cessed by theGaiaData Processing and Analysis Consortium. Support included the Bolyai Research Scholarship of the HAS (L.M., E.Pl.), ´UNKP-17-3 program of the Ministry of Hu- man Capacities of Hungary ( ´A.J.), and NKFIH grants PD-116175, PD-121203, K-115709.
References
Abazajian, K. N., et al.,ApJS 182, 543-558 (2009),arXiv: 0812.0649 Belokurov, V., et al.,MNRAS 466, 4711 (2017),arXiv: 1611.04614 Breiman, L., Machine Learning45, 5 (2001)
Chambers, K. C., et al., ArXiv e-prints (2016),arXiv: 1612.05560 Clementini, G., et al.,A&A595, A133 (2016),arXiv: 1609.04269 Drake, A. J., et al.,ApJ 696, 870 (2009),arXiv: 0809.1394 European Space Agency, ESA SP-1200 (1997)
Eyer, L., et al., ArXiv e-prints (2017),arXiv: 1702.03295
Gaia Collaboration, et al.,A&A595, A1 (2016a),arXiv: 1609.04153 Gaia Collaboration, et al.,A&A595, A2 (2016b),arXiv: 1609.04172 Guyon, I., Elisseeff, A., J. Machine Learning Res.3, 1157 (2003) Hernitschek, N., et al.,ApJ 817, 73 (2016),arXiv: 1511.05527 Iorio, G., et al.,MNRAS 474, 2142 (2018),arXiv: 1707.03833 Ivezi´c, ˇZ., et al.,AJ 120, 963 (2000),arXiv: astro-ph/0004130 Perryman, M. A. C., et al.,A&A323(1997)
Pojmanski, G.,Acta Astron.47, 467 (1997) Stokes, G. H., et al.,Icarus148, 21 (2000)
Udalski, A., Szyma´nski, M. K., Szyma´nski, G., Acta Astron. 65, 1 (2015), arXiv: 1504.05966
Wo´zniak, P. R., et al.,AJ 127, 2436 (2004),arXiv: astro-ph/0401217
